Finally, a question on this sub that is not about an “AI girlfriend” (ahem RP)
There are about a dozen + different ways to incorporate KGs into an LLM workflow with our without RAG. Some examples:
## Analyze user question, map it into KG nodes and extract connectivity links between them. Then put that info into the LLM prompt to better guide the answer.
Example: “Who is Mary Lee Pfeiffer’s son and what is he known for”? (b.t.w. try this on ChatGPT 3.5)
KG contribution – resolve Mary Lee Pfeiffer, use “gave-birth-to” edge / link to resolve Tom Cruise
Add this info to the user prompt, have LLM complete the rest of the background info, like movies appeared in, etc.
## Use KG for better RAG relevancy.
Example: Assume your KG is not about concepts but simply links paragraphs/chunks together. This could be simple as mining links like (see Paragraph X for more detail), Doing semantic similarity between chunks, putting in structural info like (chunk is part of Chapter X, Page Y), topic or concept -based connectivity between chunks.
Then, given a user query, find the most relevant starting chunk, Apply logic for what is “more relevant” from your application to figure out which other linked chunks to pull into the context. One simple hack, using node centrality or Personalized PageRank is to pull in chunks that are indirectly connected, but have high prominence in the graph
Thank you for you answer!
I’ve worked hard to improve my personal RAG implementation, searching (and asking here) ad nauseam to find ways to enhance the performance of the retrivial process…
i will study over this approach linked in the OP post, and your answer really helped me to take everything to a more “practical / tangibile” level.
I’ll try to integrate that on my experimental pipeline (currently I’m stable on RAG fusion using “query expansion” and hybrid search using transformer, SPLADE and bm25.
i already tried an approach that need a LLM to iterate over every chunk before generating embedding, mainly to solve pronouns and cross reference between chunks… Good results… But not good enough if analyzed in relation to the resource needed to iterate the llm over every item.
Maybe the integration of this “knowledge nodes/edges generation” in my “llm” pre processing will change the pro/cons balance since, from a rapid test, the model seem able to do both text preprocessing and concept extraction in the same run.
Thanks again!
.
Finally, a question on this sub that is not about an “AI girlfriend” (ahem RP)
I had many good discussions on this sub, and I really like that community… Anyway, i got your point Lol.
Thanks for this. I’ve only worked with RAG on OpenAI models and there’s a lot of prompt finetuning needed to get decent results. A KG helps define the semantic elements and relationships between document fragments and the user query for RAG.
That said, I’m still relying on the vector database to do most of the heavy lifting of filtering relevant results before feeding them into an LLM. Having an LLM clean up or summarize the user query and create a KG from the vector database’s response could lead to more accurate answers.
Having an LLM clean up or summarize the user query and create a KG from the vector database’s response could lead to more accurate answers.
That is the promise. Of course, you still need to figure out for your app domain if doing a concept-level, chunk level, or some in-between option like CSKG is the right application.
One thing I find helpful with prompt design is to spend less attention on writing instructions, replacing them with specific examples instead. This replaces word-smithing with in-context learning samples. You build up the examples iteratively, running the same prompt through more text, fixing it and adding onto the example list… until you reach your context budget for the system prompt.
Yeah, that’s what I do too. Example input and JSON key output, for example. The example idea also works with calculations: instead of telling the LLM each calculation step, use real numbers and show the result of each step in sequence.
Sometimes vector search gets inaccurate results with really short queries, those with misspellings or SMS-speak. I find it helps to get an LLM to expand and correct a query before creating an embedding vector out of it.
That’s really interesting, thank for sharing!!
How do the querying process work for this 'knowledge graph"?
Finally, a question on this sub that is not about an “AI girlfriend” (ahem RP)
There are about a dozen + different ways to incorporate KGs into an LLM workflow with our without RAG. Some examples:
## Analyze user question, map it into KG nodes and extract connectivity links between them. Then put that info into the LLM prompt to better guide the answer.
Example: “Who is Mary Lee Pfeiffer’s son and what is he known for”? (b.t.w. try this on ChatGPT 3.5)
## Use KG for better RAG relevancy.
Example: Assume your KG is not about concepts but simply links paragraphs/chunks together. This could be simple as mining links like (see Paragraph X for more detail), Doing semantic similarity between chunks, putting in structural info like (chunk is part of Chapter X, Page Y), topic or concept -based connectivity between chunks.
Then, given a user query, find the most relevant starting chunk, Apply logic for what is “more relevant” from your application to figure out which other linked chunks to pull into the context. One simple hack, using node centrality or Personalized PageRank is to pull in chunks that are indirectly connected, but have high prominence in the graph
Thank you for you answer! I’ve worked hard to improve my personal RAG implementation, searching (and asking here) ad nauseam to find ways to enhance the performance of the retrivial process…
i will study over this approach linked in the OP post, and your answer really helped me to take everything to a more “practical / tangibile” level.
I’ll try to integrate that on my experimental pipeline (currently I’m stable on RAG fusion using “query expansion” and hybrid search using transformer, SPLADE and bm25.
i already tried an approach that need a LLM to iterate over every chunk before generating embedding, mainly to solve pronouns and cross reference between chunks… Good results… But not good enough if analyzed in relation to the resource needed to iterate the llm over every item. Maybe the integration of this “knowledge nodes/edges generation” in my “llm” pre processing will change the pro/cons balance since, from a rapid test, the model seem able to do both text preprocessing and concept extraction in the same run.
Thanks again!
.
I had many good discussions on this sub, and I really like that community… Anyway, i got your point Lol.
Thanks for this. I’ve only worked with RAG on OpenAI models and there’s a lot of prompt finetuning needed to get decent results. A KG helps define the semantic elements and relationships between document fragments and the user query for RAG.
That said, I’m still relying on the vector database to do most of the heavy lifting of filtering relevant results before feeding them into an LLM. Having an LLM clean up or summarize the user query and create a KG from the vector database’s response could lead to more accurate answers.
That is the promise. Of course, you still need to figure out for your app domain if doing a concept-level, chunk level, or some in-between option like CSKG is the right application.
One thing I find helpful with prompt design is to spend less attention on writing instructions, replacing them with specific examples instead. This replaces word-smithing with in-context learning samples. You build up the examples iteratively, running the same prompt through more text, fixing it and adding onto the example list… until you reach your context budget for the system prompt.
Yeah, that’s what I do too. Example input and JSON key output, for example. The example idea also works with calculations: instead of telling the LLM each calculation step, use real numbers and show the result of each step in sequence.
Sometimes vector search gets inaccurate results with really short queries, those with misspellings or SMS-speak. I find it helps to get an LLM to expand and correct a query before creating an embedding vector out of it.